Local-first · Source-cited · Falsifiable
Anne turns model-tagged work into deterministic, source-cited memory. It captures locally, resolves semantic addresses with inspectable code, catches stale knowledge against ground truth, and abstains when the evidence is not there.
Private pilot build · Access by invitation
What changed in connected-agent startup?
Anne Desktop 1.0.8 uses a bounded, current-workspace-first startup card and makes Anne the first continuity query.
Controlled comparison: the same judge recovered stale claims at F1 0.913 with repository ground truth versus 0.667 from the memory stream alone.
Two-minute product demo
A live-corpus walkthrough of deterministic addressing, private source drill-back, reconciliation, distillation, and federated verification.
After semantic tagging
A model may help resolve what a passage is. Once an attribute is resolved, Anne uses ordinary code to decide whether it has one exact address inside the relevant structural hub. The model does not get to reinterpret that address on every retrieval.
Model-assisted consensus tags sit at the ingest boundary.
Anne measures document frequency inside the structural hub.
One attribute, one Context Sort, one source timeframe.
If the address is not unique, broader retrieval handles only what remains.
Honest boundary: semantic tagging and free-form query-to-attribute resolution remain fallible. The deterministic claim begins after the attribute is resolved; non-unique cases abstain or fall back.
The missing layer
Memory products optimize what comes back. Anne also asks whether it should come back—and whether the source can still defend it.
Claude Code, Codex, and Gemini CLI can share one encrypted corpus without sending the corpus to a memory vendor.
Anne returns compact working context with drill-back anchors to the conversation, artifact, or manifest behind each claim.
Remembered claims are checked against authoritative ground truth so completed, reversed, and stale work does not silently return.
When Anne cannot produce the evidence, it says so. The boundary between known and guessed is part of the product.
One working loop
Anne treats chat as ground truth and everything else as a map over it. That map is continuously refined, reconciled, and delivered back to the next agent at the moment it matters.
Knowledge distillation
Organizations do not only lose facts. They lose their tacit success model: what good looks like, what to watch for, and how you know it is working. Anne selects that load-bearing judgment from AI-assisted work, places it at the project, team, or organization level where it belongs, and keeps a drill-back path to the work that earned it.
Keep the non-obvious decision, reversal, signal, and success criterion.
A project lesson should not silently become an organization-wide rule.
Cross-person and independent-source support raises a claim; one voice does not.
The compact principle is useful because the exact source remains available locally.
Anne injects a bounded workspace card at startup. When an agent needs prior decisions, artifacts, deployments, or “what happened last time,” it queries Anne before repeating broad disk or web archaeology—then checks current source and tests before acting. That is where the token savings compound.
Supporting result, not the whole thesis: a frontier query-blind distiller retained near-raw answerability at about 300 tokens versus an 836-token source window (3.33 vs 3.42/5, n=12). The tested local 8B did not. Selection precision, altitude placement, and cross-person corroboration remain the product gates under active validation.
The network layer
Anne begins as verified memory inside one user or organization. Federation lets private nodes strengthen shared knowledge without pooling their corpora. Source-side verification is the proof protocol that keeps that exchange honest.
Manuals, service history, and internal engineering records.
Raw source stays localField outcomes, confirmed fixes, and local operating knowledge.
Raw source stays localNodes exchange bounded receipts while evidence remains with its custodian.
The proposed fact can be tested without exposing private chat or source prose.
A claim stays provisional until independent source families support it.
Commercial beachhead
Anne’s first vertical product turns an equipment manufacturer’s service knowledge into a verified troubleshooting assistant for its support desk and dealer network.
Check the controller’s input mode before replacing the receiver.
A founder-native market where troubleshooting still lives in tribal knowledge, PDF manuals, and the heads of senior technicians.
Equipment manufacturers with measurable warranty pain and dealer networks that amplify every verified fix.
Each deployed Anne can become a private verification node, compounding trusted field knowledge without centralizing the corpus.
Founder
“I spent twenty years troubleshooting show systems under pressure. Then I watched AI agents lose decisions the same way field teams lose tribal knowledge. Anne is the system I wanted in both rooms.”
Design partners
Anne is seeking equipment manufacturers willing to test verified, source-cited troubleshooting with a real support team and real documents.